گروه‌بندی دوندگان بر اساس سینماتیک سه‌بعدی مچ پا با استفاده از الگوریتم خوشه‌بندی عمیق زمانی

Background and Aims: The annual incidence rate of running-related injuries ranges from 19% to 79%, with about 70% affecting the knee and lower leg regions. Foot type and functional movement are crucial in assessing lower limb function, with certain foot types being predisposing factors for injuries. Previous studies have focused on static foot classification methods to predict foot movement function, but a weak correlation has been reported. To address this limitation, researchers are now exploring dynamic foot classification methods that reflect foot movement during activities like running.

Recent research has looked at identifying movement patterns in homogeneous groups using unsupervised clustering algorithms. However, these algorithms have limitations when applied to biomechanical data due to their design for static data with small dimensions. Principal component analysis is commonly used to capture data variance, but it may overlook subtle differences in movement patterns. A more effective approach is the Deep Temporal Clustering (DTC) algorithm, a deep learning method that combines dimensionality reduction and temporal clustering in a single framework. This algorithm optimizes clustering and dimensionality reduction objectives simultaneously, making it suitable for analyzing time series data.

The aim of this study was to use the DTC algorithm to group runners based on three-dimensional kinematic patterns of the ankle joint during running. We hypothesized that (1) identifying distinct subgroups based on ankle joint kinematics during the stance phase of running would be feasible and (2) there are differences in kinematic patterns among the identified groups.

Materials and Methods: 108 healthy participants (55 males and 53 females) ran barefoot at a speed of 3 meters per second. Kinematic data were collected using six Simi Motion cameras from Germany, recording at 200 Hz. Three successful attempts were recorded for each participant. The stabilization phase of running was determined using vertical force reaction data from a force plate. Raw data were filtered with a fourth-order Butterworth low-pass filter at a cutoff frequency of 16 Hz, determined through residual analysis. Ankle joint angles were calculated in MATLAB using the Cardan sequence method and normalized to 100% of the stabilization phase. Data was then transferred to Python for further analysis.

In Python 3.8, the DTC algorithm was applied to cluster kinematic patterns. The process involved three stages: dimension reduction and learning short-scale waveforms with a convolutional neural network, learning temporal relationships with a Bi-directional Long Short-Term Memory (Bi-LSTM), and non-parametric clustering on the hidden representations from Bi-LSTM to identify spatio-temporal dimensions and create clusters. The clusters were compared using one-dimensional parametric statistical mapping.

Results: Three distinct clusters were identified, with a silhouette index value of 0.74 indicating optimal clustering. Cluster 1 had the highest percentage of individuals (54%), compared to clusters 2 (24%) and 3 (26%). Individuals in cluster 1 exhibited larger average ankle dorsiflexion between 40% and 80% of the stance phase compared to the other clusters. Cluster 3 showed greater variation in ankle dorsiflexion between 60% and 100% of the stance phase. However, changes in ankle joint angles in the horizontal plane during running were similar across all three groups.

Conclusion: The main goal of this study was to use a deep temporal clustering algorithm to categorize healthy runners based on their ankle joint movement patterns while running. The identification of three distinct clusters confirms the hypothesis that there are homogeneous subgroups with unique ankle joint movement patterns during the stance phase of running. These differences in movement patterns among the clusters could help identify individuals at risk of injury and inform the development of targeted preventive or therapeutic interventions.

The second hypothesis of the research was confirmed through statistical analysis, showing significant differences in movement patterns on the sagittal and frontal planes among the identified groups. Group 1 showed a greater range of motion in the sagittal plane from 40% to 80% of the stance phase compared to the other groups, while group 3 exhibited more eversion in the final 40% of the stance phase compared to groups 1 and 2. This eversion pattern in group 3 is consistent with previous findings by Oliver et al. (2022), who identified a pronation foot group characterized by increased eversion during the stance phase. Prolonged ankle pronation, indicated by increased eversion during running, is associated with injuries such as Achilles tendonitis and stress fractures. Recognizing the pronation foot movement pattern in healthy runners may be linked to their ability to dynamically control foot pronation. Research suggests that foot muscles are less activated in a static state, highlighting the importance of assessing foot anatomy under dynamic conditions to understand individual pronation control. Static measurements may only reflect the response of inactive foot structures, such as bone alignment and tendon support, to specific static loads.

Foot muscles play a crucial role in controlling pronation and altering the foot arch shape when bearing 50% to 100% of body weight, affecting compensatory mechanisms during activities like walking. However, the limited volume of foot muscles restricts their compensatory capacity. In this study, the peak vertical reaction force during the stance phase of running exceeded twice the body weight, surpassing the foot muscles' control over pronation and arch shape. This finding is consistent with Kelly et al.'s research, suggesting differences in running loads and the foot muscles' inability to compensate for ankle joint pronation during running. Dynamic classification of runners' foot type can identify injury-prone individuals based on ankle joint kinematics, supporting existing literature on running movement patterns. This study underscores the shift from static foot classification to dynamic assessment for understanding foot movement performance during activities.

Dynamic foot type clustering through an unsupervised approach provides significant advantages over static classification methods. The model evaluates and clusters runners' movement performance to recommend targeted preventive and therapeutic interventions, including appropriate sports shoes. By identifying similar movement patterns within subgroups using a deep temporal clustering algorithm, this model reduces individual variability in response to interventions and provides optimal guidance for functional clustering and intervention prescription.